Norman Carver

Minimizing communication cost in a distributed Bayesian network using a decentralized MDP

In complex distributed applications, a problem is often decomposed into a set of subproblems that are distributed to multiple agents. We formulate this class of problems with a two layer Bayesian Network. Instead of merely providing a statistical view, we propose a satisficing approach to predict the minimum expected communication needed to reach a desired solution quality. The problem is modelled with a decentralized MDP, and two approximate algorithms are developed to find the near optimal communication strategy for a given problem structure and a required solution quality.

A Multi-Agent Architecture for Distributed Domain-Specific Information Integration

On both the public Internet and private Intranets, there is a vast amount of data available that is owned and maintained by different organizations, distributed all around the world. These data resources are rich and recent; however, information gathering and knowledge discovery from them, in a particular knowledge domain, confronts major difficulties. The objective of this article is to introduce an autonomous methodology to provide for domain-specific information gathering and integration from multiple distributed sources.

Performance evaluation of DPS coordination strategies modelled in pi-calculus

Distributed problem solving (DPS) is the subfield of multi-agent systems concerned with using systems of agents to solve large-scale, distributed problems like data interpretation in sensor networks. Coordination of agent actions is a key issue in DPS. There are not yet methods that can automatically produce effective coordination strategies for most real-world applications. We envision a tool that would support human engineering by allowing a strategy to be modelled at various levels of abstraction, with incomplete specification of the inter- and intra-agent order of actions. The tool would be able to analyse various properties of such strategies, determine the best possible time performance, and derive ordering constraints to guarantee best performance. This paper reports on research to develop key elements for such a tool, including using pi-calculus as a formal framework for defining DPS coordination strategies, and techniques for evaluating the time performance of such strategies defined with pi-calculus.

Deadlock verification of a DPS coordination strategy and its alternative model in pi-calculus

A key issue for distributed problem solving (DPS) systems is coordination of the agents actions, and methods for producing effective coordination strategies remain an active area of research. Because there are not yet approaches that can automatically produce such strategies, some human engineering is often still necessary. As a result, there is a need for a formal tool to support such human engineering. In a previous work (Khorasani et al., 2009), we investigated the use of pi-calculus as a formal language for modelling DPS coordination strategies and showed how such models could be used to evaluate the time performance of a strategy. In this paper, we focus on verification of coordination strategies. More specifically, we utilise the formal semantics of pi-calculus to detect deadlocks in a coordination strategy. We also show how, by imposing certain constraints on the pi-calculus model, one would be able to design a deadlock-free coordination strategy.

Analyzing the efficiency of strategies for MAS-based sensor interpretation and diagnosis

One of the factors holding back the application of multi-agent, distributed approaches to large-scale sensor interpretation and diagnosis problems is the lack of good techniques for predicting the performance of potential systems. In this paper we use a consideration of Bayesian network inference algorithms to construct formulas that describe the computational and communication resources required by several strategies for MAS-based distributed SI/diagnosis.